On June 10, 2026, DeepSeek released DeepSeek V4, an open-weight mixture-of-experts model with 1.6 trillion total parameters and 32 billion active parameters per token. It ships under the MIT license with a 128,000 token context window. The model appeared on DeepSeek’s official homepage and became the top trending open model on Hugging Face within 72 hours. The same week, Zyphra, Nous Research, and Cohere followed with open releases aimed at local reasoning, agentic tool calling, and enterprise retrieval. This cluster marks the clearest June startup wave in open-source AI. If you need a full ranked list, see our best open-source LLM models for 2026 coding and local agentic benchmarks.

DeepSeek published DeepSeek V4 as an open-weight model, not a closed API preview. You can download it from Hugging Face and run it with vLLM, SGLang, or llama.cpp. Zyphra released Zaya1, an 8 billion parameter sparse reasoning model under Apache 2.0. Nous Research shipped Hermes Agent, a tool-calling model built on Llama 4 Scout. Cohere released Command A+, an open-weight multilingual model for retrieval augmented generation. Each vendor posted model cards on its own homepage, and Hugging Face mirrored the weights. This is a rare month where multiple startups shipped weights before any paid tier announcement.

Why does this matter for free AI users? Closed providers are tightening access. Google cut Gemini free tier compute quotas, Anthropic replaced flat-rate Claude access with credit pools, and GitHub Copilot moved to usage-based billing. Open-weight models offer a direct workaround. You pay once for hardware or use free Hugging Face Inference endpoints. The license terms also matter. MIT and Apache allow commercial use without royalty. CC-BY-NC and similar research licenses prevent paid products. Before adopting, check the specific license. For context, our AI free tier limits get tougher in June 2026 covers the paid API squeeze that makes these open models valuable.

The June 2026 startup release pattern is led by smaller labs, not only megacorps. DeepSeek and Zyphra are venture-backed startups. Nous Research is an independent research collective. Cohere is a platform company increasingly open-weight. Their releases target distinct workloads: large-scale reasoning, on-device mobile deployment, autonomous agents, and enterprise search. The result is a diversified open ecosystem. In the rest of this article, we compare the four standout June 2026 startup releases. Each item includes model specs, license, benchmark scores where announced, and how to run it locally.

How Do the Top Options Compare?

Release Best For License Context Window Key Spec
DeepSeek V4 Large-scale reasoning and coding MIT 128,000 tokens 1.6T total parameters, 32B active MoE
Zyphra Zaya1 8B On-device reasoning and mobile Apache 2.0 32,000 tokens 8B MoE, matches 70B dense on math
Nous Hermes Agent Autonomous tool calling and agents Apache 2.0 131,000 tokens Built on Llama 4 Scout, Hermes function calling
Cohere Command A+ Enterprise RAG and multilingual CC-BY-NC 4.0 128,000 tokens 111B total parameters, open-weight

Specs reflect vendor announcements as of June 2026. License terms can change between model cards and fine-tunes. Always verify the exact model card on Hugging Face or the vendor homepage.

1. DeepSeek V4 , Best for Large-Scale Reasoning and Coding

DeepSeek V4 is the biggest open-weight startup release of June 2026. DeepSeek shipped the model on June 10, 2026 under an MIT license. It uses a mixture-of-experts design with 1.6 trillion total parameters and 32 billion active parameters per token. The context window is 128,000 tokens. That context length supports long codebases, full research papers, and multi-turn agent logs. You can download the weights from Hugging Face or start from DeepSeek’s official homepage at DeepSeek. The release was not gated. No phone verification, no waitlist, no usage quota.

The model’s MIT license means you can fine-tune, merge, or ship it in a commercial product without royalty obligations. The tradeoff is size. You need about 800GB of VRAM for full FP16 inference. 4-bit quantization brings that down to around 200GB, still beyond a single consumer GPU. Most developers will use the smaller llama.cpp GGUF quants or cloud GPU rentals. If you want the full technical breakdown, read our DeepSeek V4 open-source 2026 coverage.

Benchmark claims are strong. DeepSeek reported 89.1 on MMLU-Pro, 92.4 on GPQA Diamond, and top-three results on LiveCodeBench. Those numbers put V4 within striking distance of closed frontier models. The company did not release RLHF preference weights, so instruction following relies on community fine-tunes. That is the main limitation compared to ChatGPT or Claude. You get raw capability, not a polished product.

Key strengths:

  • ✅ 1.6T total parameters with 32B active delivers frontier-class reasoning
  • ✅ MIT license allows commercial fine-tuning and distribution
  • ✅ 128K context handles long code and document tasks
  • ✅ Street benchmark scores place it near closed models
  • ✅ No per-token API cost once self-hosted
  • ❌ Requires high-end hardware for full precision, around 800GB VRAM
  • ❌ No first-party chat interface or preference-tuned weights
  • ❌ Large disk footprint for full checkpoints

Who it’s for: Developers with access to an A100/H100 node who want frontier-like open weights without vendor lock-in.

2. Zyphra Zaya1 8B , Best for On-Device Reasoning

Zyphra released Zaya1 on June 12, 2026. It is an 8 billion parameter mixture-of-experts reasoning model under Apache 2.0. The context window is 32,000 tokens. Zyphra positioned the model for edge devices, phones, and laptops. The sparse architecture means only 1.8 billion parameters are active per token, which keeps latency low. You can pull weights from the Zyphra homepage or Hugging Face. The company also released ONNX and Core ML exports for local app integration.

Zaya1 focuses on math and code reasoning. Zyphra said it matches a 70 billion dense model on GSM8K and MATH while using far less memory. That makes it a practical option for local AI assistants that need to solve problems without cloud calls. The Apache 2.0 license grants patent rights and permits commercial use. The main drawback is the 32K context, which is short for long agent sessions. You can find detailed runtime guidance in our Zaya1 8B Zyphra open-source reasoning 2026 article.

Hardware requirements are modest. You can run the 4-bit quantized version in about 6GB of RAM on a MacBook. Full FP16 needs around 16GB. That is a stark contrast to DeepSeek V4. Zyphra did not release a chat-tuned version with RLHF. The base model needs prompting or fine-tuning for conversational use.

Key strengths:

  • ✅ 8B MoE with only 1.8B active runs on laptops and phones
  • ✅ Apache 2.0 license includes patent grants and commercial use
  • ✅ Strong math and code benchmarks for the size
  • ✅ ONNX and Core ML exports simplify edge deployment
  • ✅ Modest memory use, roughly 6GB at 4-bit
  • ❌ 32K context window limits long agent or document tasks
  • ❌ No RLHF chat-tuned weights in the first release
  • ❌ Reasoning focus is narrower than general instruction models

Who it’s for: Mobile and edge developers who need private reasoning without cloud costs.

3. Nous Hermes Agent , Best for Autonomous Tool Calling and Agents

Nous Research released Hermes Agent on June 17, 2026. The model is built on Llama 4 Scout and fine-tuned for tool calling, API usage, and multi-step agent workflows. It ships under Apache 2.0 with a 131,000 token context window. Nous published the model on Nous Research’s homepage and Hugging Face. It is a drop-in replacement for systems that currently call GPT-4o or Claude via function calling. The model supports parallel tool calls, structured JSON output, and long-horizon planning.

The practical advantage is that Hermes Agent handles agentic tasks without paid API credits. Closed agent products changed pricing in June 2026. Anthropic replaced flat-rate access with credit pools, and GitHub Copilot moved to usage-based billing. Hermes Agent offers a self-hosted alternative. You can run it with llama.cpp, vLLM, or Ollama. The 131K context window is enough for agent memory and tool logs. See our Hermes Agent Nous Research open-source 2026 piece for setup details.

Benchmarks are mixed. Hermes Agent scores well on Berkeley Function Calling Leaderboard and ToolBench, but it trails DeepSeek V4 on general reasoning. The model inherits Llama 4 Scout’s multilingual weakness. It works best in English and Spanish, with degraded performance in low-resource languages. Fine-tuning the base on non-English tasks is possible because the license is Apache 2.0. The main risk is that the underlying Llama 4 Scout has a different license for its base weights, so check the fine-tune card before commercial deployment.

Key strengths:

  • ✅ Purpose-built for tool calling and agentic workflows
  • ✅ 131K context window supports long agent memory
  • ✅ Apache 2.0 license on the fine-tuned weights
  • ✅ Runs on consumer GPUs via 4-bit quantization
  • ✅ Strong function calling scores on public leaderboards
  • ❌ General reasoning trails larger models like DeepSeek V4
  • ❌ Multilingual performance is weaker in low-resource languages
  • ❌ Base model license terms require review before commercial use

Who it’s for: Self-hosters building agents that need reliable tool calls without per-token API pricing.

4. Cohere Command A+ , Best for Enterprise RAG and Multilingual Search

Cohere released Command A+ on June 20, 2026. It is an open-weight model with 111 billion total parameters and a 128,000 token context window. Cohere targeted the release at retrieval augmented generation, enterprise search, and multilingual business workflows. The model is available from Cohere’s homepage and mirrored on Hugging Face. It supports 23 languages and includes native tool use for connectors and databases.

The license is CC-BY-NC 4.0. That means non-commercial use is free, but you cannot sell a hosted API or enterprise product without a separate commercial agreement from Cohere. This is common for Cohere’s open releases. The license restricts startups that want to build paid SaaS on top of Command A+. However, internal enterprise deployments may be allowed if no external monetization occurs. Read the exact terms before you ship. Our Cohere Command A Plus open-source 2026 article explains the licensing implications.

Benchmark claims are strong for retrieval. Cohere reported 86.4 on MMLU, 73.2 on multilingual MMLU, and 91.0 on the finance subset. The model is not as strong as DeepSeek V4 on code generation, but it outperforms many dense models on document QA. Hardware requirements are lower than DeepSeek V4. A 4-bit quantized version runs on a single RTX 4090 with 24GB VRAM. Full precision needs about 220GB of VRAM.

Key strengths:

  • ✅ 111B parameters with 128K context supports large document RAG
  • ✅ 23 languages with strong multilingual retrieval scores
  • ✅ Runs on a single RTX 4090 at 4-bit quantization
  • ✅ Native tool use for databases and connectors
  • ✅ Good enterprise search metrics on finance and legal benchmarks
  • ❌ CC-BY-NC 4.0 license blocks commercial hosted products without a deal
  • ❌ Code generation trails DeepSeek V4
  • ❌ Open-weight but not fully open source under strict definitions

Who it’s for: Enterprise teams that need multilingual RAG and can accept a non-commercial license or negotiate with Cohere.

Frequently Asked Questions

What is the biggest open-source AI release in June 2026?

DeepSeek V4 is the biggest by parameter count. It has 1.6 trillion total parameters and 32 billion active parameters per token under an MIT license. It supports a 128,000 token context window and is available on Hugging Face.

Which June 2026 open model is best for local on-device AI?

Zyphra Zaya1 8B is the best for on-device use. It is an 8 billion parameter mixture-of-experts model under Apache 2.0. It runs in about 6GB of RAM at 4-bit quantization and focuses on math and code reasoning.

Are these June 2026 startup models free for commercial use?

Not all. DeepSeek V4 and Zyphra Zaya1 use MIT and Apache 2.0, which allow commercial use. Cohere Command A+ uses CC-BY-NC 4.0, which prevents commercial hosted products without a separate agreement. Always check the model card.

How do these open models compare to closed models like GPT-5 or Claude?

They approach closed models on many benchmarks. DeepSeek V4 reported top-three results on LiveCodeBench and GPQA. The main gap is preference tuning and polished chat interfaces. Closed models remain easier to use out of the box.

What hardware do I need to run DeepSeek V4 locally?

Full FP16 inference needs about 800GB of VRAM. 4-bit quantization lowers that to around 200GB. Most users rent an A100 or H100 node or use a community GGUF quant on llama.cpp.

Where can I download these models?

The easiest source is Hugging Face, where all four models are mirrored. You can also visit DeepSeek, Zyphra, Nous Research, and Cohere homepages for official model cards and download links. Avoid unofficial third-party repos.

What Should You Remember?

  • DeepSeek V4 leads June 2026 with a 1.6T parameter MIT-licensed MoE model.
  • Zyphra Zaya1 brings 8B sparse reasoning to phones and laptops under Apache 2.0.
  • Nous Hermes Agent targets self-hosted tool calling with a 131K context window.
  • Cohere Command A+ serves enterprise RAG but uses a non-commercial CC-BY-NC license.
  • License checks matter: MIT and Apache allow commercial use, CC-BY-NC does not.
  • Local hardware varies from 6GB for Zaya1 to 200GB+ for DeepSeek V4 at 4-bit.
  • Closed API pressure makes June open releases a real cost escape hatch.

Free AI News is an independent editorial publication. Information about AI pricing, free-tier limits, and features changes frequently and may become outdated. Always verify current details through the vendor’s official pages. Affiliate links may earn a commission at no cost to you, and never affect our reporting.